China publishes its official energy statistics annually in the China Energy Statistical Yearbook released by the National Bureau of Statistics (NBS), including both national and provincial energy balance sheets for each province. The national energy balance sheets are revised each time an economic census is completed, and the revisions are published in the next Energy Statistical Yearbook. The China Energy Statistical Yearbooks from 2005 (CESY2005), 2009 (CESY2009) and 2014 (CESY2014) contain the revised national energy data for the periods of 1999–2003, 1996–2007 and 2000–2012, respectively, based on the results of the first, second and third national economic censuses conducted during this time. The IEA also publishes energy statistics for China, which have been widely used in international emission inventories (such as EDGAR). The IEA also regularly revises its energy statistics and is now operating in cooperation with the NBS, who annually provides the IEA with China's energy statistics, and in recent years the IEA statistics have been found to be quite consistent with the NBS's national energy balance sheet.

Six data sets of energy statistics were involved in this study: the original edition of the national energy balance sheets from the CESY (CT-CESY-Ori) and its revisions following the first economic census (CT-CESY-1C), the second economic census (CT-CESY-2C) and the third economic census (CT-CESY-3C); the provincial energy balance sheets from the CESY (PBP-CESY); and the 2012 edition of China's energy statistics from the IEA (CT-IEA-2012). These data sets are summarized in Table 1. Note that here CT-CESY-Ori represents the first edition of national energy statistics covering the whole period 1990–2013. For revised national energy statistics (i.e., CT-CESY-1C, CT-CESY-2C, CT-CESY-3C), the data were taken from the previous edition for years for which revised data were unavailable. Although energy statistics for 2014 are already published, we did not include the year 2014, for the reason that the emission inventory is being updated. The IEA energy statistics were used for comparison, but they were excluded from the uncertainty calculations in the current work. The IEA energy statistics are generally based on NBS's national energy balance sheets and currently quite consistent with CT-CESY-2C, as shown in Fig. 1. They may soon be updated based on CT-CESY-3C.

The MEIC emission inventory model (http://www.meicmodel.org) was used in this study to investigate the emission responses to different energy statistics. MEIC is a dynamic technology-based inventory developed for China covering the years from 1990 to 2013 by Tsinghua University following the work of INTEX-B (Zhang et al., 2009), with several updates, such as a unit-based emission inventory of power plants (F. Liu et al., 2015), a high-resolution vehicle emission inventory at the county level (Zheng et al., 2014) and an improved non-methane volatile organic carbon (NMVOC) speciation approach for various chemical mechanisms (Li et al., 2014). The MEIC inventory includes recent control policies based on the available official reports (Ministry of Environmental Protection of China (MEP), 1991–2014, 2000–2014). The MEIC version 1.1 (MEIC v1.1) uses energy consumption data from PBP-CESY, excluding diesel and gasoline consumption data, which are taken from the national energy statistics (currently CT-CESY-1C) because the diesel consumption data provided in the national energy statistics were thought to possibly be more reliable (Zhang et al., 2007). The emissions in MEIC were estimated as a product of the activity rate (such as energy consumption or material production), the technology distributions of fuel/production and emission control, the unabated emission factor and the removal efficiency. Thus, the emission estimates can be simplified as the activity rates multiplied by their respective net emission factors of different fuel/product types in different sectors. Note that the net emission factors in MEIC change dynamically driven by the technology renewal process year by year. Technology distributions within each sector are obtained from Chinese statistics, a wide range of unpublished statistics by various industrial association and technology reports. For example, technology distributions in the power sector were obtained from a unit-base database (F. Liu et al., 2015). Technology distributions in the transportation sector were estimated based on a fleet model (Zheng et al., 2014). The methods of emission estimates have been documented in our previous work (Zhang et al., 2007, 2009; Zheng et al., 2014; F. Liu et al., 2015).

To further explore the impact of energy data inconsistencies on estimates of China's emissions, five emission inventories based on five sets of energy statistics (i.e., CT-CESY-Ori, CT-CESY-1C, CT-CESY-2C, CT-CESY-3C and PBP-CESY) were established in the framework of the MEIC inventory. Note that only energy data were changed in the calculations of these emission inventories, while other data such as net emission factors remained the same as in the MEIC inventory. Thus the emission uncertainties derived from these inventories are only those associated with energy uncertainties. They do not include uncertainties in the emission factors and other parameters in the MEIC inventory, which is not addressed in this study. For different energy data sets, the same net emission factors were applied for fuel consumption in a given sector in each year during the emission calculations. In fact, energy differences might change the technology renewal process and further change the net emission factors. However, considering that those assumptions would likely add additional uncertainty and we do not discuss the uncertainties in emission factors, such indirect impacts on emission factors are not included in this study. We only applied all the fuel consumption differences to the combustion sectors. The sectoral categories are consistent across all the energy data sets from NBS (Table S1 in the Supplement). The same scale factor in fuel consumption was applied for all the sub-categories in the same major sector (e.g., industrial coal-fired boilers and kilns in the industrial sector, on-road diesel vehicles and off-road mobile sources in the transportation sector). The possible uncertainties in feedstocks and products resulting from energy uncertainties are not included in this study, and also the uncertainties in biomass consumption are not included due to lack of multiple data sets; thus our estimates of emission uncertainties are likely on the conservative side. As the emission calculations were performed with province-level data, energy consumption in the national energy statistics was directly allocated to provinces by using the ratios derived from the provincial energy statistics.

Apparent uncertainties in China's energy consumption for the period of 1990–2013 were quantified based on five publicly available energy statistics (i.e., CT-CESY-Ori, CT-CESY-1C, CT-CESY-2C, CT-CESY-3C and PBP-CESY), as shown in Fig. 1. Before 1996 there are no annual provincial data, and essentially just one national data set, which has not been revised. But since 1996, multiple data sets and/or revisions are available for each year. Notable apparent uncertainties have been observed since then, which can be divided into three periods: an early period (1996–2003); a more recent period of rapid growth (2004–2012); and the most recent period of convergence (2013). During the early period (1996–2003), China's energy consumption grew slowly, from 1352–1389 Mtce in 1996 to 1709–1971 Mtce in 2003. The average apparent uncertainty in total energy consumption during this period is 133 Mtce, with a peak of 261 Mtce in 2003, and the corresponding apparent uncertainty ratios are 9.0 % for the period as a whole and 14.3 % for 2003. During the recent period of rapid growth (2004–2012), along with the rapid growth in China's economy and energy consumption, the apparent uncertainty in the total energy consumption also increased, with a mean uncertainty of 449 Mtce for this period and a maximum of 646 Mtce in 2011; the corresponding apparent uncertainty ratios are 14.5 % for this period overall and 16.9 % for 2011. The inconsistencies during the early period have been reported in many previous studies (Sinton, 2001; Akimoto et al., 2006; Zhang et al., 2007), but few studies (Guan et al., 2012; Z. Liu et al., 2015; Korsbakken et al., 2016) have noted the more recent rapid-growth period. Converging uncertainties are observed in 2013, with the release of the newest energy statistics based on the third economic census – the apparent uncertainty in total energy consumption for 2013 is reduced to 62 Mtce, and the corresponding apparent uncertainty ratio is only 1.5 %. We notice that the apparent uncertainty for 2014 (not shown here) is similar to that for 2013, also much smaller than that during the recent period of rapid growth (2004–2012).

With regard to different types of energy, coal dominates the apparent uncertainties in total energy consumption. The average apparent uncertainties in coal consumption for 1996–2003, 2004–2012 and 2013 are 147, 428 and 194 Mtce, respectively, and the corresponding apparent uncertainty ratios are 14.2, 19.4 and 6.7 %. The sum of the provincial data (PBP-CESY) is generally higher than the national total (i.e., CT-CESY-Ori, CT-CESY-1C, CT-CESY-2C and CT-CESY-3C) with regard to total energy consumption and coal consumption. After each of the three economic censuses, the national total energy consumption data (CT-CESY-1C, CT-CESY-2C and CT-CESY-3C) were revised upward to approach the provincial totals, primarily by adjusting the coal-related data. The apparent uncertainties in oil consumption during 1996–2003 are relatively large, with a mean of 48 Mtce and an average apparent uncertainty ratio of 15.8 %. The provincial total oil consumption is lower than the national total for 1996–2003, but this situation is reversed between 2005 and 2011. The apparent uncertainties in the consumption of natural gas and other types of energy are smaller than the uncertainties in coal and oil, suggesting that the statistical data for natural gas and other energy sources may be more accurate because their use is generally metered.

The apparent uncertainties in coal consumption were further analyzed by sector, as shown in Fig. 2. As the largest consumer of coal in China, the power sector is found to exhibit less uncertainty in its coal consumption than other sectors. Coal consumption in the industrial sector is highly uncertain, with an apparent uncertainty ratio for 2012 of 45.4 %, which represents the greatest contribution to the total uncertainty in coal consumption. A significant decrease in coal consumption in the industrial sector during 1996–2002 is observed in the CT-CESY-Ori data, and this decrease resulted in a slight decrease in the total coal consumption. For the heating sector and the residential sector, although the levels of coal consumption in these two sectors are smaller than those in the power and industrial sectors, comparable apparent uncertainties are also found; the apparent uncertainty ratios for the heating and residential sectors in 2012 are 37.8 and 46.9 %, respectively.

To evaluate the effects of uncertainties in the energy statistics on China's emission estimates, five emission inventories based on different sets of energy statistics (i.e., CT-CESY-Ori, CT-CESY-1C, CT-CESY-2C, CT-CESY-3C and PBP-CESY) were established. Figure 3 shows the apparent uncertainties in China's emissions during 1990–2013. Figure 4 shows the apparent uncertainty ratio in China's emissions during 1990–2013. It should be noted that the emission uncertainties discussed below, which were derived from these five emission inventories, are based only on uncertainties in the energy data; thus, they could reflect the impacts of energy uncertainties on emission estimates. For the early period (1996–2003), the average apparent uncertainties for SO2, NOx, VOC, PM2.5 and CO2 are 2.10, 0.83, 0.41, 0.34 and 278 Tg, respectively, and the corresponding apparent uncertainty ratios are 10.2, 6.7, 3.2, 2.8 and 6.7 %. For the recent period of rapid growth (2004–2012), the apparent uncertainties are increasing over time and are more significant than those in the early period, although this fact has rarely been discussed in the literature; the average apparent uncertainties during this period for SO2, NOx, VOC, PM2.5 and CO2 are 5.77, 2.98, 1.60, 0.80 and 1026 Tg, respectively, and the corresponding apparent uncertainty ratios are 20.4, 12.6, 7.7, 6.4 and 12.4 %. For 2012, the apparent uncertainties for these species are 7.76, 4.68, 1.90, 1.10 and 1633 Tg, respectively, and the corresponding apparent uncertainty ratios are 30.0, 16.4, 7.7, 9.2 and 15.6 %. The apparent uncertainty for CO2 in 2010 is 1283 Tg in this study, which is similar to the discrepancy (∼ 1400 Tg) reported by Guan et al. (2012) but lower than the uncertainty in 2012. In the most recent period of convergence (2013), the apparent uncertainty ratio in emissions is less than 5 % for most species because of the lower apparent uncertainties in the energy statistics after the third economic census. Note that the emission discrepancies calculated from the provincial and national energy statistics get smaller after the third economic census (i.e., PBP-CESY and CT-CESY-3C), compared with those before the third economic census (e.g., PBP-CESY and CT-CESY-2C). For example, CO2 emission discrepancy in 2010 between PBP-CESY and CT-CESY-3C is only 548 Tg, much less than that reported by Guan et al. (2012), in which the NBS's data before the third economic census were used.

It should be noted that the apparent uncertainties calculated in this study are subjected to the energy data sets used. For example, the small apparent uncertainties before 1996 might become larger if a new energy data set that revises the data of this period is included. Apparent uncertainties during the recent period of rapid growth (2004–2012) are higher than during the early period (1996–2003), implying that the discrepancies might be accumulated and expanded for a period of rapid growth. For example, underestimates of the growth trends of small enterprises might result in accumulated underestimations. Note that the energy consumption apparently became more consistent between provincial and national statistics after the three economic censuses, indicating that the new energy statistics after the economic census may include evolved methodologies for data collection and more cross-checks to reduce the discrepancies. In this case, conducting censuses in some interval years could help to reduce the accumulated discrepancies. The apparent uncertainty ratio in the years when economic censuses were newly conducted (i.e., 2004, 2008 and 2013) is generally less than that in previous years (i.e., 2003, 2007 and 2012), as shown in Fig. 4. The converging uncertainties in 2013 may also be caused by the third economic census.

As seen, the energy-induced uncertainties in emissions differ by species; the largest uncertainties are observed for SO2, followed by NOx and CO2, and the smallest are found for PM2.5 and VOC. Taking the year 2012 as a case in which the uncertainties are prominent, the emission uncertainties were separated by sector and by energy type, as shown in Table 2. SO2 emissions are more sensitive to energy uncertainties than are CO2 emissions because of the high contribution (approximately 50 %) from industrial coal combustion, which is the largest source of uncertainty in SO2 emissions (6.04 Tg). A large fraction (approximately 24 %) of NOx emissions is contributed by the use of diesel in the transportation sector; the corresponding activity data have a lower uncertainty ratio than that for coal use, leading to a lower sensitivity than that of SO2. PM2.5 and VOC emissions also show less sensitivity to energy uncertainties because they represent relatively small contributions from energy consumption and high contributions (approximately 40–60 %) from industrial process emissions. Note that non-combustion emission uncertainty is not addressed in this study. With regard to contributions by sector, industry is the dominant sector, accounting for 77.8, 72.3, 46.8, 52.4 and 73.2 % of the total apparent uncertainties in SO2, NOx, PM2.5, VOC and CO2 emissions, respectively. Although the power sector is a major source of emissions of many species (contributing approximately 25–35 % of the total emissions of CO2, NOx and SO2), it is estimated to contribute less than 7 % to the total apparent uncertainties for all species because of the relatively low uncertainty for coal consumption in the power sector. Transportation is another key contributor to emission uncertainties for NOx (contributing 17.3 % of the uncertainty for this species), whereas the residential sector is significant for SO2 (contributing 18.3 % of the uncertainty). With regard to energy type, 97.6 and 93.8 % of the emission uncertainties of SO2 and PM2.5, respectively, originate from coal, whereas 31.2 % of the VOC emission uncertainties come from oil. The contributions of gas and other fuels are negligible because uncertainties in biomass consumption are not included and other emissions are relatively small. Note that biomass consumption, which is usually thought to be quite uncertain, would contribute more uncertainties in emissions.

Discrepancies in energy data affect not only the absolute emission estimates for individual years but also multi-year emission trends because of the interannual variability of these discrepancies. Table 3 compares the emission trends for China derived from different energy statistics. For the early period (1996–2003), slower growth rates of CO2, NOx and SO2 emissions are found from the CT-CESY-Ori inventory (22.9, 38.0 and 14.3 %, respectively) than from the PBP-CESY inventory (35.5, 47.5 and 28.0 %, respectively), which is consistent with previous studies (Akimoto et al., 2006; Zhang et al., 2007). The trends derived from the national energy statistics were revised upward after each of the three economy censuses, bringing them closer to those indicated by the provincial energy statistics. SO2 and CO2 show a dip according to the CT-CESY-Ori inventory, but this effect is not significant in the PBP-CESY inventory. SO2 emissions declined by 13.7 % during the period of 1996–2000 according to the CT-CESY-Ori inventory but increased by 1.4 % according to the PBP-CESY inventory. These differences reflect the large uncertainties in industrial coal consumption during 1996–2000 – a decline of 28.4 % is indicated by CT-CESY-Ori, whereas only a slight decrease of 3.8 % is found from PBP-CESY. It should be noted that the dip for NOx emissions is not as distinct as that for CO2. This is because the fuel consumption in the power and transportation sectors, for which the NOx emission factors are the largest, was steadily increasing during this period.

In the recent period of rapid growth (2004–2012), CO2 and NOx emissions, respectively, increased by 91.8 and 77.6 % according to the PBP-CESY inventory but by only 70.8 and 48.4 % as determined from the CT-CESY-Ori inventory; SO2 emissions increased by 1.6 % according to the PBP-CESY inventory but decreased by 18.8 % as indicated by the CT-CESY-Ori inventory. For the period from 1996 to 2012, the CO2 growth rates inferred from the CT-CESY-Ori, CT-CESY-3C and PBP-CESY inventories are 145, 172 and 191 %, respectively, similar to the growth rates in the total energy consumption (160, 197 and 207 %); the differences between different energy statistics demonstrate that trends in CO2 emissions are good indicators of trends in energy consumption. NOx and SO2 also show marked differences in emission growth – increases by 197 and 45.1 % during 1996–2012 according to the provincial energy statistics but by only 139 and 7.2 % as determined from the original national energy statistics. From 2012 to 2013, the total energy consumption and CO2 emissions, respectively, increased by 3.7 and 3.7 % as seen from CT-CESY-3C but decreased by 0.4 and 2.1 % according to PBP-CESY. The GDP increased by 7.7 % between 2012 and 2013; thus, the decreasing trend in CO2 emissions indicated by PBP-CESY is unexpected. Korsbakken et al. (2016) also pointed out that initial claims that Chinese CO2 emissions fell in 2014 according to preliminary official statistics were probably premature. The unexpected energy and CO2 emission decline in 2013 in PBP-CESY could be explained by the fact that the PBP-CESY data for 2013, which probably include updates based on the third economic census, are closer to the data from CT-CESY-3C. As a result, the total growth rates since 1996 indicated by PBP-CESY and CT-CESY-3C are more similar to each other for the period 1996–2013 than periods before 2013 (e.g., 1996–2012). As part of the Chinese five-year plan, the Chinese government established a set of targets for emission reduction, including a 10 % SO2 reduction from the 2005 levels by 2010 and reductions of 10 % in NOx and 8 % in SO2 from the 2010 levels by 2015. Our results show that, because uncertainties in energy statistics can lead to the inference of different emission trends, reliable energy data are crucially important for obtaining accurate estimates of both the absolute levels of emissions and their trends.

The large uncertainties between the national and provincial energy statistics can be explained in terms of both inadequacies in China's statistical system and artificial factors. First, China's statistical data are generally collected and reported from bottom to top, and there is a lack of effective means of cross-checking at the local level; thus, these data are faced with problems such as data inconsistency and double counting (Wang et al., 2014). Inconsistencies between interprovincial imports and exports have been found in the provincial energy statistics – the sum of coal “interprovincial imports” is higher than the sum of coal “interprovincial exports” (Zhang et al., 2007). Also, provincial statistics more likely include double counting because certain interprovincial activities are claimed by all provinces involved. A similar situation affects economic statistics: the aggregate provincial GDP in 2012 was approximately 11 % larger than the national total. Second, unlike for large- and medium-size enterprises, which have defined data collection and reporting procedures, the energy data for small enterprises are merely estimated, which could strongly degrade the data quality of the energy statistics (Jiang et al., 2009; Wang and William, 2011; Wang, 2014). A typical example is that the original official statistics (CT-CESY-Ori) did not fully count the coal production of illegal small coal mines, leading to underestimations in coal production around 2000 (Wang and William, 2011; Guan et al., 2012). Third, certain signs indicate that energy data may be modified for artificial purposes (Guan et al., 2012). The energy revisions after the second economic census (CT-CESY-2C) were found to bring the country closer to achieving its energy conservation targets (Aden, 2010). We also notice that some provinces had zero statistical difference; i.e., the supply data matches the consumption data exactly, which might mean that some provincial data were adjusted to achieve the exact match.

We compared China's coal consumption in 1996–2013 as indicated by different energy statistics from the supply perspective, as shown in Fig. 5. In the supply approach, energy consumption is estimated based on production, trade and changes in stock (consumption = production - exports + imports + change in stock + statistical difference). From the supply perspective, the national energy statistics tend to estimate conservative production and thus underestimate coal consumption. The coal consumption data for 1996–2012 from the national energy statistics were revised upward after the first (CT-CESY-1C), second (CT-CESY-2C) and third (CT-CESY-3C) censuses because of increasing coal production, which may be largely explained by the small coal mines that were initially unaccounted for in the official statistics (Guan et al., 2012). Meanwhile, inconsistencies in interprovincial transport manifest as interprovincial net imports (see Fig. 5), resulting in a higher coal supply in the provincial energy statistics, implying that either coal production is underestimated or coal consumption is overestimated. For the years before 2008, the coal production indicated by the provincial energy statistics is reasonably consistent with that derived from the original national energy statistics (CT-CESY-Ori) and lower than that from the revised national energy statistics (CT-CESY-3C), which could help partially explain the interprovincial net imports during this period as underestimates in production. Moreover, the provincial energy statistics likely include double counting and thus might result in overestimates. This effect may be more significant in recent years with the more frequent collaboration among companies at the provincial level.

Satellite observations, which have been widely used in the assessment of emission trends in previous studies (e.g., Richter et al., 2005; van de A et al., 2008; Stavrakou et al., 2008; Lamsal et al., 2011), could be used as one independent approach to verifying energy statistics. Akimoto et al. (2006) compared trends in bottom-up NOx emissions with satellite-derived NO2 columns for the period of 1996–2002 and found that the emission trends derived from various energy statistics were all lower than those inferred from the satellite observations (which increased by 50 %). The PBP-CESY trends were within the uncertainty of the satellite observations, whereas the IEA2004 and CT-CESY trends were apparently underestimated beyond the uncertainty of the satellite observations. Zhang et al. (2007) compared trends over China in bottom-up NOx emissions with satellite NO2 columns observed from 1996 to 2004 and found a larger trend in the satellite NO2 columns than in the NOx emissions. Berezin et al. (2013) derived top-down estimates of CO2 emission trends by means of the satellite-derived NOx emission trends obtained using an inverse model and CO2-to-NOx emission ratios (i.e., CO2 / NOx) from bottom-up inventories. They also found a significant quantitative difference between bottom-up and indirect top-down estimates of the CO2 emission trend for the period of 1996–2001, and the difference for the period of 2001–2008 was found to be in the range of possible systematic uncertainties associated with their estimation method.

Previous studies have investigated the decline in energy consumption between 1997 and 2001 indicated by CT-CESY-Ori (Akimoto et al., 2006; Sinton, 2001; Sinton and Fridley, 2000; Zhang et al., 2007; Berezin et al., 2013). This supposed decline was completely eliminated after the revisions following the three censuses. This fact may support the conclusion of these early studies that the trend in energy consumption indicated by the provincial energy statistics was more accurate than that derived from the unrevised national energy statistics during the early period. Z. Liu et al. (2015) estimated total Chinese energy consumption by adopting the apparent consumption approach and estimated a value for 2000–2012 that was 10 % higher than that reported in China's national statistics before the third economic census and lower than that from the provincial energy statistics. The discrepancies between the national and provincial energy statistics were reduced after the three economic censuses. These facts indicate that the newest energy statistics after the third economic census may include evolved methodologies for data collection and more cross-checks to reduce inconsistencies between the national and provincial energy statistics and thus can be recommended for use. In contrast, the energy consumption indicated by the national energy statistics from before the third economic census may be underestimated because of underestimations in energy production, whereas the energy consumption indicated by the provincial energy statistics from before the third economic census may be overestimated because of double counting.

In this study, we find that uncertainties in energy statistics have great impacts on China's emission estimates, which could also be used to partially explain different emission estimates from other inventories. The Ministry of Environmental Protection of China tends to estimate lower SO2 and NOx emissions than MEIC (e.g., 30 % lower for SO2 and 20 % lower for NOx in 2012). Lower energy consumption from the national energy statistics, compared with provincial energy statistics, could help to explain the differences in emissions. Trends in CO2 emissions are good indicators of trends in energy consumption, which can reflect the differences in energy statistics between different inventories. We compared the emission trends for CO2 and NOx in MEIC, the Asian inventory REAS and the global inventory EDGAR (Fig. 6). From 1996 to 2001, CO2 emissions increased by 9.6 and 9.2 % according to MEIC and REAS, respectively, but increased by only 0.4 % according to EDGAR; total energy consumption increased by 11.5 and 3.1 % according to PBP-CESY and CT-CESY-1C, respectively, during the same period. From 1996 to 2008, CO2 emissions increased by 135 % according to MEIC but increased by only 114 % according to EDGAR; total energy consumption increased by 138 and 110 % according to PBP-CESY and CT-CESY-1C, respectively. These differences indicate that the energy consumption indicated by EDGAR, which was created using the IEA energy statistics, is likely closer to the national energy statistics.

The differences in NOx emission trends could be partially explained by differences in the energy statistics. During the period of 1996–2008, NOx emissions increased by 127 % according to MEIC but by only 76 % according to EDGAR. If the CO2 growth trend were to be replaced with that from MEIC while keeping the same NOx-to-CO2 ratios, a greater (93 %) increase in NOx emissions would be found from EDGAR. However, significant differences also arise from the emission factors (i.e., the NOx-to-CO2 ratios): the overall NOx-to-CO2 ratio decreased by only 3.5 % in MEIC for the 1996–2008 period but decreased by 17.9 % in EDGAR. Similar trends in the overall NOx-to-CO2 ratio are found between MEIC and REAS: it increased from 1996 to 2001, primarily driven by a faster growth rate of diesel consumption (46–51 %), which has a higher NOx-to-CO2 ratio, than of coal consumption (-15 to 7 %), but it then decreased after 2004, primarily because of the implementation of NOx emission standards in the power and transportation sectors. It should be noted that EDGAR tends to estimate a much earlier and more rapid decline in NOx emission factors compared with those seen from MEIC and REAS (see Fig. 6c), for which the underlying driving forces are difficult to understand. For example, the NOx-to-CO2 ratios in EDGAR began to decrease significantly for the power and transportation sectors in 1993 and 1990, respectively (see Fig. 6d), earlier than the years of implementation of major control measures regarding NOx emissions in these sectors (1996 and 2001, respectively) (State Environmental Protection Administration of China (SEPA), 1996, 2001).

The CO2-to-NOx emission ratios taken from bottom-up inventories could be an important potential source of error in top-down estimates of CO2 emission trends based on satellite NO2 columns. Berezin et al. (2013) used the emission ratios from EDGAR and found an increase in CO2 emissions of as high as 240 % for the 1996–2008 period using the top-down approach, much larger than the trends observed in bottom-up inventories (e.g., 114 % in EDGAR). These substantial differences should be attributable mainly to the rapidly increasing CO2-to-NOx ratios in EDGAR. If we adopt the emission ratios from MEIC (including uncertainties), we find an increase of 147–197 %, much closer to the values from bottom-up inventories. Although uncertainties still exist, these results indicate that the energy consumption from EDGAR, which is similar to CT-CESY-1C, as well as energy consumption in 2008 from CT-CESY-2C, is likely to be underestimated.

Top-down estimates of the CO2-to-NOx emission ratios using satellite observations could offer an alternative approach. Reuter et al. (2014) used top-down estimation methods and found that the CO2-to-NOx emission ratio for the years 2003–2011 in east Asia had increased by 4.2 ± 1.7 % yr-1. They found a large positive trend in CO2 emissions in east Asia (9.8 ± 1.7 % yr-1) that exceeded the positive trend in NOx emissions (5.8 ± 0.9 % yr-1). The MEIC inventory reports a similar CO2 trend in China (10.4 % yr-1) during the same period. Reuter et al. (2014) noted a considerably smaller CO2 trend in EDGAR (6.9 % yr-1) than in these top-down estimates; it appears that considering the possible underestimations in Chinese CO2 trends in EDGAR due to uncertainties in energy statistics could help to explain this difference. The MEIC inventory reports a larger NOx trend in China (8.1 % yr-1) than that reported by Reuter et al. (2014) for east Asia, which is consistent with Wang et al. (2014), who also found a faster NOx growth rate in China (34 %) than in east Asia as a whole (25 %) for 2005–2010.

Zhao et al. (2011) estimated the uncertainties (i.e., 95 % confidence intervals around the central estimates) of Chinese total SO2, NOx and PM2.5 emissions in 2005 to be -14 to 13, -13 to 37 % and -17 to 54 %, respectively. We found that the apparent uncertainty ratios arising from the 2012 energy statistics for SO2 and NOx emissions could be as large as 30.0 and 16.4 %, respectively, indicating the importance of energy statistics to Chinese emission estimates for recent years, especially for SO2 and NOx. Variations in energy consumption could be an important source of emission uncertainties for SO2 and NOx. For VOC and PM2.5, uncertainties in energy consumption act as a minor source due to emission contributions from non-energy activities and large uncertainties from emission factors.